MATHEMATICAL AND COMPUTATIONAL ANALYSIS OF YOLOV11 AND EFFICIENTNET-B0 FOR PLANT DISEASE CLASSIFICATION

This paper compares two deep-learning models for recognizing plant diseases from leaf images: the EfficientNet-B0 classifier and the YOLOv11 detector, both tested on the PlantVillage dataset. Instead of only reporting accuracy numbers, we show the mathematics behind them. We write out the main formulas each model uses — image normalization, the depthwise-separable convolution, the SiLU activation, softmax cross-entropy, intersection-over-union, and the CIoU, distribution-focal, and cross-entropy losses of the detector — and then work through the numbers by hand. For example, one 3×3 layer is shown to need 6.55 times fewer operations when depthwise separation is used, a cross-entropy of 0.417 is computed for a sample output, an intersection-over-union of 0.681 gives a CIoU loss of 0.328, and a distribution-focal loss of 0.773 is found. We also rebuild the precision, recall, and F1 scores from confusion counts (for one class: precision 99.0%, recall 98.0%, F1 98.5%), compute an average precision of 0.943, and run a paired t-test that gives t(4)=5.62, p≈0.005. EfficientNet-B0 reaches 98.5% accuracy with 5.3 M parameters and 0.39 GMACs, while YOLOv11 reaches 97.8% accuracy and a mAP@0.5 of 96.2% but needs about 55 times more computation. The classifier is faster and slightly more accurate, while the detector is more expensive but also locates the disease.